On probabilistic representation of non-probabilistic belief revision (Q1118576)

The belief-states of an ideally rational agent can be modelled either by a ``belief-set'', i.e. a consistent set of propositions closed under logical deduction, or by a probability function P. The addition of new evidence x leads to a ``belief-revision''. In the former framework, such a revision amounts to transforming the earlier belief-set A into a revised belief-set \(A\oplus x\), whereas revision in terms of probability- functions basically amounts to conditionalization P*x. Both approaches are related to each other since, in particular, one may associate to each probability function P the ``top of P'', i.e. the set of all propositions x such that \(P(x)=1\), as a belief-set. The ``non-uniqueness''-problem arises from the fact that, conversely, to every belief-set A there may exist different probability-functions \(P_ 1,P_ 2,..\). which have A as their ``top''; moreover, the revised functions \(P_ 1*x,P_ 2*x,..\). may in general have different ``tops'' so that they do not uniquely determine a revised belief-set \(A\oplus x\) as ``the'' top of \(P_ 1*x\) or \(P_ 2*x\).... In this paper five strategies for solving this problem are discussed. The first 3 approaches (one being based on the assumption of ``top equivalence''; the second consisting in choosing the ``common part'' of the non-equivalent ``tops''; the third one relying on ``base- functions'') are all rejected as inadequate. The authors suggest to adopt either a ``relational approach'' or to take the notion of a ``belief- state'' as a primitive one which allows to redefine belief-sets and belief-function in such a way as to avoid the non-uniqueness problem.